FW-DBNs: Feedback-Weighted Dynamic Bayesian Networks for Real-Time Velocity Prediction

Accurate velocity prediction is a practically crucial yet challenging task for intelligent vehicles. The challenge derives from the time-varying vehicle states, and the dynamic environment. Dynamic Bayesian Networks (DBNs) are widely used to predict vehicle velocity in dynamic environments due to their inherent capability to handle uncertainty and the flexibility to incorporate expert prior knowledge for constructing network structures. However, discrete-to-continuous state transitions restrict the outputs of traditional DBNs to discrete levels, thus being intractable for real-time applications. Moreover, the first-order Markov assumption fails to adequately utilize historical information, leading to poor long-term prediction performance. To this end, a feedback-weighted DBNs (FW-DBNs) framework is proposed for predicting the vehicle velocity in dynamic environments. First, we proposed an adaptive window search algorithm leveraging historical data to narrow down the search domain to a smaller one, thus expediting the search process. Second, we integrated multiple DBNs with various prediction horizons, forming a feedback-weighted mechanism, which allows adaptive adjustments to the weighted values. The ablation experiment results demonstrate that the adaptive window search algorithm reduces the running time by 47.6%, while the feedback-weighted mechanism improves prediction performance by over 20%. Experimental results across various scenes on the highD dataset demonstrate that FW-DBNs outperformed the state-of-the-art methods. The driver-hardware-in-the-loop simulation across four typical scenarios validated the efficiency and robust generalization ability of the FW-DBNs.